Thirty-Five Genomic Compressors Put To The Test In Massive Study
New benchmark reveals deep trade-offs between file size, memory use, and processing speed across traditional and AI-driven genetic compressors.
Thirty-five different genetic data compression programs have been put to the test in a major new evaluation, revealing surprising trade-offs in how the scientific community shrinks massive genomic files. As sequencing costs plummet and genetic datasets swell into the tens of thousands of terabytes across international archives, managing storage space has become a critical bottleneck for modern biology laboratories.
Long-read sequencing technologies from platforms like Oxford Nanopore Technologies and PacBio generate massive files containing reads exceeding 10,000 base pairs, overcoming many limitations of older short-read methods. However, storing and transmitting these sprawling sequence archives demands highly efficient lossless compression algorithms that can shrink file sizes without dropping a single nucleotide of biological data.
How Thirty-Five Compressors Stack Up Against Each Other
To understand which software performs best, the authors of the new study built LRCBench, a comprehensive evaluation suite designed to test reference-free lossless compressors. The evaluation, available on the arXiv preprint server, systematically analyzed 35 distinct algorithms using 16 real-world datasets spanning multiple species, read lengths, and sequencing platforms.
The tested software fell into four distinct categories: traditional general-purpose tools, learning-based general-purpose methods, traditional dedicated genomic compressors, and learning-based dedicated algorithms. The researchers measured each tool across multiple dimensions, including compression ratio, storage savings percentage, total processing runtime, peak memory consumption, and robustness against varying data distributions.
Which Compressors Achieved the Smallest File Sizes
The results indicate that dedicated genomic compressors generally achieve stronger compression ratios than standard general-purpose tools. Among all 35 tested programs, a tool named CoLoRd secured the best overall compression ratio, averaging 1.050 bits per base and delivering a storage saving percentage of nearly 87 percent across the full benchmark suite. Other dedicated systems, such as the GeCo series, also placed near the top for raw storage reduction.
However, achieving maximum compression comes with heavy computational costs. CoLoRd required over four hours of processing time and more than 8 gigabytes of peak memory during operation. By contrast, lightweight general-purpose utilities offered massive advantages in speed and memory efficiency, though their file sizes remained noticeably larger.
The Heavy Computational Toll of Artificial Intelligence Models
The evaluation also tested emerging artificial intelligence and deep learning compressors, though their high computational demands restricted most tests to smaller MinION datasets. While advanced neural architectures like CMIX achieved exceptional compression ratios by mixing the probability distributions of over 2,000 internal models, the computational price was astronomical.
Running CMIX required more than 400 combined hours for compression and decompression on a subset of files. Other learning-based tools encountered severe runtime bottlenecks or memory overhead spikes, highlighting a wide performance gap between theoretical neural compression efficiency and practical usability in a standard laboratory setting.
What Lies Ahead for Genomic Data Management
The authors suggest that future genomic compressor design must tackle several key hurdles, including hardware parallel acceleration using graphics processing units, privacy-sensitive security frameworks to protect human genetic data, and automated hyper-parameter tuning to simplify complex software for non-computer scientists. As sequencing technologies continue to evolve, these benchmarks provide essential guidance for researchers striving to balance storage budgets with computational speed.
The study was posted to the arXiv preprint server on November 18, 2026.
This article has been fact checked for accuracy, with information verified against reputable sources. Learn more about us and our editorial process.
Last reviewed on .
Article history
- Latest version
Reference(s)
- Ma, Huidong. “Benchmarking Traditional and Learning-Based Reference-Free Compressors for Long-Read Sequencing Data.” arXiv.org <http://arxiv.org/abs/2611.23794v1>.
Cite this page:
- Posted by Elizabeth Taylor